Implementation method of convolutional neural network module for enhancing channel rearrangement and fusion

Abstract

The invention discloses an implementation method of a convolutional neural network module for enhancing channel rearrangement and fusion. The method comprises the following steps: extracting features of a feature map of a certain level through a convolutional layer, and generating a transitional feature map; carrying out pooling on the generated transition feature map in a spatial dimension to obtain a rearrangement fusion vector with the same size as a channel; and calculating a new rearranged and fused channel and a corresponding weight by using the obtained rearranged and fused vector, accumulating a feature map of a certain channel to the new channel and an adjacent channel according to the calculated weight, traversing all channels, and obtaining a fused feature map after all channels are subjected to weighted accumulation. The module provided by the invention can be seamlessly inserted into any convolutional neural network, enhances communication of information of different channels, and can be applied to a network of image classification and a backbone network of tasks such as target detection, semantic segmentation and the like.

Description

technical field [0001] The invention belongs to the field of artificial intelligence, relates to machine learning and deep learning, and specifically relates to an implementation method of a convolutional neural network module that strengthens channel rearrangement and fusion. Background technique [0002] Deep convolutional neural networks have achieved great success in computer vision, subverting most of the traditional computer vision algorithms. A deep convolutional network usually consists of convolution (Convolution, referred to as Conv), batch normalization (BatchNormalization, referred to as BN), and linear rectification function (Rectified Linear Unit, referred to as ReLU) as a basic computing unit, Conv / BN / ReLU. Most convolutional neural networks use this basic unit as an element to connect and stack in different ways. In the stacking process, convolution layers with different convolution kernel sizes are usually used to achieve different functions. In particular,...

AI Technical Summary

Problems solved by technology

[0005] The present invention proposes a neural network plug-and-play module that adjusts channel arrangement and strengthens channel fusion, which can be used to improve the problem of insufficient fusion of different grouped channels after depth-separable convolution or grouped convolution, and enrich the single-channel The amount of information, in the process of practical application, even non-group convolution can be used

Images